A physiology-guided multimodal learning framework for EEG-free driver behavior recognition using peripheral biosignals

Accurate driver behavior recognition is important for driver monitoring systems and intelligent transportation applications. Although EEG-based approaches can provide useful neural information, their practical use is often limited by sensing complexity, preparation burden, and sensitivity to motion artifacts. This study proposes a physiology-guided multimodal learning framework for EEG-free driver behavior recognition using peripheral biosignals. A publicly available multimodal driving physiology dataset was used, including recordings from 35 participants performing five driving behaviors: smooth driving, acceleration, deceleration, lane change, and turning. Three peripheral physiological signals, namely ECG, EMG, and GSR, were used as model inputs. The proposed framework incorporates physiological knowledge into multimodal representation learning through modality-aware encoders tailored to the distinct characteristics of ECG, EMG, and GSR. A physiology-guided fusion mechanism is further introduced to model complementary cross-modal interactions among the three biosignals. Experimental results showed that the proposed ECG+EMG+GSR model achieved an average accuracy of 0.782 and a macro-F1 score of 0.783 under a five-fold evaluation protocol with separated training, validation, and test sets, with MCC and Cohen’s kappa reaching 0.711 and 0.711, respectively. Compared with EEG-only and full-modality baselines reported in the original dataset study, the proposed framework demonstrated the feasibility of EEG-free driver behavior recognition using only peripheral biosignals. Under the matched ECG+EMG+GSR input setting, the proposed framework outperformed the strongest baseline across all four evaluation metrics. In addition, sensing-configuration analysis showed that ECG+EMG+GSR provided a favorable balance between recognition performance and sensing burden. These results indicate that physiology-guided multimodal learning can effectively exploit heterogeneous peripheral biosignals for accurate and interpretable EEG-free driver behavior recognition.